189 lines
5.7 KiB
Python
189 lines
5.7 KiB
Python
import importlib.metadata
|
|
import math
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
|
|
try:
|
|
from flash_attn.flash_attn_interface import flash_attn_varlen_func
|
|
except ImportError:
|
|
flash_attn_varlen_func = None
|
|
try:
|
|
from sageattention import sageattn_varlen
|
|
@torch.compiler.disable()
|
|
def sageattn_varlen_func(
|
|
q,
|
|
k,
|
|
v,
|
|
cu_seqlens_q,
|
|
cu_seqlens_kv,
|
|
max_seqlen_q,
|
|
max_seqlen_kv,
|
|
):
|
|
return sageattn_varlen(q, k, v, cu_seqlens_q, cu_seqlens_kv, max_seqlen_q, max_seqlen_kv)
|
|
except:
|
|
pass
|
|
|
|
MEMORY_LAYOUT = {
|
|
"flash_attn": (
|
|
lambda x: x.view(x.shape[0] * x.shape[1], *x.shape[2:]),
|
|
lambda x: x,
|
|
),
|
|
"sageattn_varlen": (
|
|
lambda x: x.view(x.shape[0] * x.shape[1], *x.shape[2:]),
|
|
lambda x: x,
|
|
),
|
|
"sdpa": (
|
|
lambda x: x.transpose(1, 2),
|
|
lambda x: x.transpose(1, 2),
|
|
),
|
|
"sageattn": (
|
|
lambda x: x.transpose(1, 2),
|
|
lambda x: x.transpose(1, 2),
|
|
),
|
|
"vanilla": (
|
|
lambda x: x.transpose(1, 2),
|
|
lambda x: x.transpose(1, 2),
|
|
),
|
|
}
|
|
|
|
|
|
def get_cu_seqlens(text_mask, img_len):
|
|
"""Calculate cu_seqlens_q, cu_seqlens_kv using text_mask and img_len
|
|
|
|
Args:
|
|
text_mask (torch.Tensor): the mask of text
|
|
img_len (int): the length of image
|
|
|
|
Returns:
|
|
torch.Tensor: the calculated cu_seqlens for flash attention
|
|
"""
|
|
batch_size = text_mask.shape[0]
|
|
text_len = text_mask.sum(dim=1)
|
|
max_len = text_mask.shape[1] + img_len
|
|
|
|
cu_seqlens = torch.zeros([2 * batch_size + 1], dtype=torch.int32, device="cuda")
|
|
|
|
for i in range(batch_size):
|
|
s = text_len[i] + img_len
|
|
s1 = i * max_len + s
|
|
s2 = (i + 1) * max_len
|
|
cu_seqlens[2 * i + 1] = s1
|
|
cu_seqlens[2 * i + 2] = s2
|
|
|
|
return cu_seqlens
|
|
|
|
|
|
def attention(
|
|
q,
|
|
k,
|
|
v,
|
|
mode="flash_attn",
|
|
drop_rate=0,
|
|
attn_mask=None,
|
|
causal=False,
|
|
cu_seqlens_q=None,
|
|
cu_seqlens_kv=None,
|
|
max_seqlen_q=None,
|
|
max_seqlen_kv=None,
|
|
batch_size=1,
|
|
):
|
|
"""
|
|
Perform QKV self attention.
|
|
|
|
Args:
|
|
q (torch.Tensor): Query tensor with shape [b, s, a, d], where a is the number of heads.
|
|
k (torch.Tensor): Key tensor with shape [b, s1, a, d]
|
|
v (torch.Tensor): Value tensor with shape [b, s1, a, d]
|
|
mode (str): Attention mode. Choose from 'self_flash', 'cross_flash', 'torch', and 'vanilla'.
|
|
drop_rate (float): Dropout rate in attention map. (default: 0)
|
|
attn_mask (torch.Tensor): Attention mask with shape [b, s1] (cross_attn), or [b, a, s, s1] (torch or vanilla).
|
|
(default: None)
|
|
causal (bool): Whether to use causal attention. (default: False)
|
|
cu_seqlens_q (torch.Tensor): dtype torch.int32. The cumulative sequence lengths of the sequences in the batch,
|
|
used to index into q.
|
|
cu_seqlens_kv (torch.Tensor): dtype torch.int32. The cumulative sequence lengths of the sequences in the batch,
|
|
used to index into kv.
|
|
max_seqlen_q (int): The maximum sequence length in the batch of q.
|
|
max_seqlen_kv (int): The maximum sequence length in the batch of k and v.
|
|
|
|
Returns:
|
|
torch.Tensor: Output tensor after self attention with shape [b, s, ad]
|
|
"""
|
|
pre_attn_layout, post_attn_layout = MEMORY_LAYOUT[mode]
|
|
q = pre_attn_layout(q)
|
|
k = pre_attn_layout(k)
|
|
v = pre_attn_layout(v)
|
|
|
|
if mode == "sdpa":
|
|
if attn_mask is not None and attn_mask.dtype != torch.bool:
|
|
attn_mask = attn_mask.to(q.dtype)
|
|
x = F.scaled_dot_product_attention(
|
|
q, k, v, attn_mask=attn_mask, dropout_p=drop_rate, is_causal=causal
|
|
)
|
|
elif mode == "sageattn_varlen":
|
|
x = sageattn_varlen_func(
|
|
q,
|
|
k,
|
|
v,
|
|
cu_seqlens_q,
|
|
cu_seqlens_kv,
|
|
max_seqlen_q,
|
|
max_seqlen_kv,
|
|
)
|
|
# x with shape [(bxs), a, d]
|
|
x = x.view(
|
|
batch_size, max_seqlen_q, x.shape[-2], x.shape[-1]
|
|
) # reshape x to [b, s, a, d]
|
|
elif mode == "flash_attn":
|
|
x = flash_attn_varlen_func(
|
|
q,
|
|
k,
|
|
v,
|
|
cu_seqlens_q,
|
|
cu_seqlens_kv,
|
|
max_seqlen_q,
|
|
max_seqlen_kv,
|
|
)
|
|
# x with shape [(bxs), a, d]
|
|
x = x.view(
|
|
batch_size, max_seqlen_q, x.shape[-2], x.shape[-1]
|
|
) # reshape x to [b, s, a, d]
|
|
elif mode == "vanilla":
|
|
scale_factor = 1 / math.sqrt(q.size(-1))
|
|
|
|
b, a, s, _ = q.shape
|
|
s1 = k.size(2)
|
|
attn_bias = torch.zeros(b, a, s, s1, dtype=q.dtype, device=q.device)
|
|
if causal:
|
|
# Only applied to self attention
|
|
assert (
|
|
attn_mask is None
|
|
), "Causal mask and attn_mask cannot be used together"
|
|
temp_mask = torch.ones(b, a, s, s, dtype=torch.bool, device=q.device).tril(
|
|
diagonal=0
|
|
)
|
|
attn_bias.masked_fill_(temp_mask.logical_not(), float("-inf"))
|
|
attn_bias.to(q.dtype)
|
|
|
|
if attn_mask is not None:
|
|
if attn_mask.dtype == torch.bool:
|
|
attn_bias.masked_fill_(attn_mask.logical_not(), float("-inf"))
|
|
else:
|
|
attn_bias += attn_mask
|
|
|
|
# TODO: Maybe force q and k to be float32 to avoid numerical overflow
|
|
attn = (q @ k.transpose(-2, -1)) * scale_factor
|
|
attn += attn_bias
|
|
attn = attn.softmax(dim=-1)
|
|
attn = torch.dropout(attn, p=drop_rate, train=True)
|
|
x = attn @ v
|
|
else:
|
|
raise NotImplementedError(f"Unsupported attention mode: {mode}")
|
|
|
|
x = post_attn_layout(x)
|
|
b, s, a, d = x.shape
|
|
out = x.reshape(b, s, -1)
|
|
return out
|